S
SUJAN
Agent Architecture · ACTIVE

Agent Memory

Exploring persistent episodic memory and state compactors for multi-turn autonomous coding agents.

TELEMETRY: 1.83s · 4 tools · 12k tokens
TOOLS: Python pgvector LangGraph Embeddings

Hypothesis & Experiment

Can we preserve long-term agent context across 500+ turns without linear token cost growth or degradation of historical reasoning?

Methodology

We test hierarchical compaction where raw tool input/outputs are summarized asynchronously into an immutable knowledge graph while retaining strict constraint keys.

Observations

  • Fact preservation improved by 41% compared to sliding-window context buffers.
  • Sub-graph queries against indexed decisions yielded sub-25ms retrieval.